Validity of the SRS in Minimally Verbal Children

Background:  Implicit in any comparison of scores between groups is the assumption that an instrument is measuring the construct in the same way across groups; when this assumption is tested and confirmed, the instrument is said to have the property of measurement invariance (MI). Despite being an essential psychometric property, MI is often overlooked. This property is especially important for a measure of ASD symptoms, given the heterogeneity in presentation of this condition.

Objectives:  The goal of this analysis was to test MI of the Social Responsiveness Scale, Second Edition across minimally verbal (MV) and verbal (V) children with ASD. Several items on the SRS-2 relate to verbal abilities, suggesting that scores for MV children should be interpreted with caution. Although considerably information on the concurrent/predictive validity of the SRS exists, neither the Total Score nor the Treatment Subscales of the SRS-2 were empirically derived, giving rise to the possibility that initial model fit may be poor and MI irrelevant. Thus, we first assessed the Total Score and the Treatment Subscales, but followed with evaluation of two empirically derived solutions: the SCI and RRB two-factor solution specified in the SRS-2 manual and the five-factor solution proposed by Frazier et al. (2014).

Methods:  Data were drawn from the Simons Simplex Collection, comprising children with nonverbal IQ less than 70 and no phrase speech (MV; n=298) and children with nonverbal IQ greater than 70 and phrase speech (V; n=2,338). Invariance was tested in a multiple group confirmatory factor analytic framework. Model fit, and change in model fit as a result of increasing parameter constraints, was assessed with several complementary statistics (only RMSEA and ∆McNCI are reported here).

Results:  Fit statistics are shown in Table 1. Neither the Total Score nor the Treatment Subscale factor models fit the data, with the exception of the Awareness factor, which had adequate fit. Configural and metric invariance were also supported for the Awareness factor, but scalar was not. The fit of both the two-factor model and the Frazier model to the data were adequate; configural and metric invariance were supported, but not scalar.

Conclusions: These analyses demonstrated first that the Total Score and Treatment Subscales generally do not represent unidimensional constructs, and we were unable to produce psychometric support for their use, regardless of language level. Although we did find evidence for the adequate fit of the Awareness Treatment Subscale, mixed evidence for the fit of the two-factor model, and confirmed the adequacy of the Frazier et al. structure, scalar invariance of any organization was not supported across the MV and V subgroups. This means SRS-2 scores reflect both symptom severity and some unmeasured factor that differs between groups; functionally, this suggests that it is inappropriate to use these subscales without accounting for verbal ability, and that scores may not be combined or compared across groups.